Blind Adaptive Interference Suppression Based on ... - Semantic Scholar

1

Blind Adaptive Interference Suppression Based on Set-Membership Constrained Constant-Modulus Algorithms with Dynamic Bounds Rodrigo C. de Lamare and Paulo S. R. Diniz Abstract— This work presents blind constrained constant modulus (CCM) adaptive algorithms based on the set-membership filtering (SMF) concept and incorporates dynamic bounds for interference suppression applications. We develop stochastic gradient and recursive least squares type algorithms based on the CCM design criterion in accordance with the specifications of the SMF concept. We also propose a blind framework that includes channel and amplitude estimators that take into account parameter estimation dependency, multiple access interference (MAI) and intersymbol interference (ISI) to address the important issue of bound specification in multiuser communications. A convergence and tracking analysis of the proposed algorithms is carried out along with the development of analytical expressions to predict their performance. Simulations for a number of scenarios of interest with a DS-CDMA system show that the proposed algorithms outperform previously reported techniques with a smaller number of parameter updates and a reduced risk of overbounding or underbounding. Keywords— Interference suppression, blind adaptive estimation, setmembership estimation, spread spectrum systems.

actual one) is significantly increased, leading to a performance degradation. In addition, when the measured noise in the system is time-varying and the multiple access interference (MAI) and the intersymbol interference (ISI) encountered by a receiver in a wireless network are highly dynamic, the selection of an error-bound is further complicated. This is especially relevant for low-complexity estimation problems encountered in applications that include mobile units and wireless sensor networks [9], [10], where the sensors have limited signal processing capabilities and power consumption is of central importance. These problems suggest the deployment of mechanisms to automatically adjust the error bound in order to guarantee good performance and a low update rate (UR). A. Prior and Related Work

I. I NTRODUCTION Set-membership filtering (SMF) [2], [3], [5], [6] is a class of recursive estimation algorithms that, on the basis of a predetermined error bound, seeks a set of parameters that yield bounded filter output errors. These algorithms have been applied to a variety of applications including adaptive equalization [5] and multi-access interference suppression [6], [7]. The SMF algorithms are able to combat conflicting requirements such as fast convergence and low misadjustment by introducing a modification on the objective function. These algorithms exhibit reduced complexity due to data-selective updates, which involve two steps: a) information evaluation and b) update of parameter estimates. If the filter update does not occur frequently and the information evaluation does not involve much computational complexity, the overall complexity can be significantly reduced. Adaptive SMF algorithms usually achieve good convergence and tracking performance due to the use of an adaptive step size or an adaptive forgetting factor for each update. This translates into reduced complexity due to the data selective updating. However, the performance of SMF techniques depends on the error-bound specification, which is very difficult to obtain in practice due to the lack of knowledge of the environment and its dynamics. In wireless networks characterized by non-stationary environments, where users often enter and exit the system, it is very difficult to choose an error bound and the risk of overbounding (when the error bound is larger than the actual one) and underbounding (when the error bound is smaller than the Part of this work has been presented at ICASSP 2006. This work was supported in part by CNPQ and FAPERJ. Dr. Rodrigo C. de Lamare is with the Communications Research Group, University of York, United Kingdom. Prof. Paulo S. R. Diniz is with LPS/COPPE-UFRJ, Rio de Janeiro, Brazil. Emails: [email protected], [email protected]

In this context, blind methods are appealing because they can alleviate the need for training sequences or pilots, thereby increasing the throughput and efficiency of wireless networks. In particular, blind algorithms based on constrained optimization techniques are important in several areas of signal processing and communications such as beamforming and interference suppression. The constrained optimization required in these applications deals with linear constraints that correspond to prior knowledge of certain parameters such as the direction of arrival (DoA) of user signals in antenna array processing [23] and the signature sequence of the desired signal in DS-CDMA systems [24], [25]. Therefore, linear signal models and constraints can be used to describe various wireless communications systems including multi-input multi-output (MIMO) and orthogonal frequency-division multiplexing (OFDM) systems. For instance, linear constraints that incorporate the knowledge of user signatures of a DS-CDMA system can also be used to exploit the knowledge of the spatial signatures of MIMO systems. A number of blind algorithms with different trade-offs between performance and complexity have been reported in the last decades [24]-[36]. Local scattering, synchronization and estimation errors, imperfectly calibrated arrays and imprecisely known wave field propagation conditions are typical sources of uncertainties that can lead to a performance degradation of blind algorithms. The literature indicates that the CCM-based algorithms [30][36] have a superior performance to algorithms based on the constrained minimum variance (CMV) criterion [24]- [27]. The CCM-based algorithms exploit prior knowledge about the constant modulus property of signals like M-PSK, which results in an improved performance over CMV-type techniques and a performance that is very close to the linear minimum mean-square error (MMSE) training-based techniques. Moreover, the CCM-

2

type algorithms are robust against errors in the effective signature sequence required for blind parameter estimation, which prevents a severe performance degradation in the presence of uncertainties. These features make CCM-type algorithms excellent candidates for interference suppression in wireless networks. The need for the adjustment of parameters (step size, forgetting factor) and the computational complexity of CCM-based techniques calls for approaches like SMF. Prior work on SMF blind algorithms for interference suppression is very limited [7], [8]. The use of time-varying bounds is also restricted to applications where one assumes that the ”true” error bound is constant [16] and to the parameter-dependent error bound recently proposed in [13], [15]. The time-varying bound techniques so far reported are not blind and do not exploit these mechanisms for channel and parameter estimation. B. Contributions In this work, we propose set-membership blind adaptive constrained algorithms based on the CCM criterion. Preliminary results have been reported in [22]. In particular, we derive stochastic gradient (SG) and recursive least squares (RLS)-type CCM algorithms designed in accordance with the specifications of the SMF concept. The second contribution is a low-complexity blind framework for parameter estimation, and tracking of parameter evolution, MAI and ISI that relies on simple estimation techniques and employs the proposed set-membership CCM algorithms with a time-varying bound. The third contribution of this work is an analysis of the optimization problem that gives rise to the proposed algorithms along with a mean-squared error (MSE) convergence and tracking analysis using the energy conservation approach [39] for predicting the performance of the algorithms. A simulation study considers an interference suppression application to DS-CDMA systems, which compares the performance of the proposed and existing algorithms, and discusses the main features of the algorithms. This paper is structured as follows. Section II describes the DS-CDMA system model and briefly reviews linearly constrained receivers. Section III introduces the SM blind algorithms with time-varying bound. Section IV proposes blind parameter dependent and interference dependent bounds along with blind channel and amplitude estimation algorithms. Section V is dedicated to the analysis of the proposed algorithms. Section VI is devoted to the presentation and discussion of numerical results, while Section VII gives the conclusions. II. DS-CDMA SYSTEM MODEL AND L INEARLY C ONSTRAINED R ECEIVERS Let us consider the uplink of a symbol synchronous DSCDMA system with K users, N chips per symbol and Lp propagation paths. A synchronous model is assumed for simplicity since it captures most of the features of more realistic asynchronous models with small to moderate delay spreads. The modulation is assumed to have constant modulus. Let us assume that the signal has been demodulated at the base station, the channel is constant during each symbol and the receiver is perfectly synchronized with the main channel path. The received signal after filtering by a chip-pulse matched filter and sampled

at chip rate yields an M -dimensional received vector at time i r[i] =

K ∑

Ak [i]bk [i]C k hk [i] + η k [i] + n[i],

(1)

k=1

where M = N + Lp − 1, n[i] = [n1 [i] . . . nM [i]]T is the complex Gaussian noise vector with zero mean and E[n[i]nH [i]] = σ 2 I independent and identically distributed samples, (.)T and (.)H denote transpose and Hermitian transpose, respectively, and E[.] stands for expected value. The user symbols are denoted by bk [i], the amplitude of user k is Ak [i], and η k [i] is the intersymbol interference (ISI) for user k. The signature of user k is represented by ck = [ck (1) . . . ck (N )]T , the M × Lp constraint matrix C k that contains one-chip shifted versions of the signature sequence for user k and the Lp × 1 vector hk [i] with the multipath components are described by   ck (1) 0     .. .. h [i] k,0   . ck (1)  .    .. Ck =   , hk [i] =  . .. .   ck (N ) .   hk,Lp −1 [i] .. . ck (N ) 0 (2) The MAI comes from the non-orthogonality between the received signature sequences. The ISI originates from the multipath propagation effects of the channel, depends on the length of the channel response and how it is related to the length of the chip sequence. We define Ls as the ISI span, i.e., the number of symbols affected by the channel. For Lp = 1, Ls = 1 (no ISI), for 1 < Lp ≤ N, Ls = 3, for N < Lp ≤ 2N, Ls = 5, and so on. At time instant i we will have ISI coming not only from the previous time instants but also from the next symbols. The linear model in (1) can be used to represent other wireless communications systems including multi-input multi-output (MIMO) and orthogonal frequency-division multiplexing (OFDM) systems. For example, the user signatures of a DS-CDMA system are equivalent to the spatial signatures of a MIMO system. In order to describe the design of linearly constrained receivers, we consider the received vector r[i], the M × Lp constraint matrix C k that contains one-chip shifted versions of the signature sequence for user k and the Lp × 1 vector hk [i] = [hk,0 [i] . . . hk,Lp −1 [i]]T with the multipath components to be estimated. The CCM linear receiver design is equivalent to determining an FIR filter wk [i] with M coefficients that provide an estimate of the desired symbol as follows zk [i] = wH k [i]r[i],

(3)

where the detected symbol is given by ˆbk [i] = Q(wH k [i]r[i]), where Q(·) is a function that performs the detection according to the constellation employed. The design of the receive filter wk [i] is based on the optimization of the CM cost function ] [ 2 2 (4) JCM (wk [i]) = E (|wH k [i]r[i]| − 1) subject to the constraints given by wH k [i]pk [i] = ν, where pk [i] = C k hk [i] is the effective signature vector that corresponds to the convolution between the original signature sequence and the channel gains, and ν is a constant to ensure the

3

convexity of the optimization problem as will be discussed later on. This approach assumes the knowledge of the channel. However, when multipath is present these parameters are unknown and time-varying, requiring channel estimation. The CCM receive filter expression that iteratively solves the constrained optimization problem in (4) is given by )−1 [ ( −1 H [i]p [i] [i] d [i] − p [i]R wk [i + 1] = R−1 k k k k k ( ) ] −1 H · pk [i]Rk [i]dk [i] − ν pk [i] , i = 1, 2, . . . (5) 2 H where zk [i] = wH k [i]r[i], Rk [i] = E[|zk [i]| r[i]r [i]], dk [i] = ∗ E[zk [i]r[i]]. A detailed derivation of the CCM estimation approach can be found in [33], [34], [36]. It should be remarked that the expression in (5) is a function of previous values of the filter wk [i] and therefore must be iterated in order to reach a solution. In addition to this, the iterative method in (5) assumes the knowledge of the channel parameters. Since there is a large number of applications that have to deal with unknown multipath propagation, it is also important to be able to blindly estimate the multipath components. In order to blindly estimate the channel, a designer can adopt the blind channel estimation procedure based on the subspace approach reported in [26], [41] and which is described by

ˆ k [i] = arg min h

hk [i]

H −1 hH [i]C k hk [i] k [i]C k R

(6)

wH adapk [i]r[i]. Since it is not practical to predict all data pairs, ∩i tive methods work with the membership set ψi = m=1 Hm provided by the observations, where Hm = {wk ∈ C M : |zk [i]| ≤ γk [i]}. In order to devise an effective SM algorithm, the bound γk [i] must be appropriately chosen. Due to the time-varying nature of many practical environments, this bound should also be adaptive and capable of estimating certain characteristics for the SM estimation technique. We devise SM-CCM algorithms equipped with variable step sizes and forgetting factors that are able to automatically tune to different situations, which is an advantage over methods operating with fixed parameters in time-varying scenarios. In what follows, we derive SM-CCM algorithms for blind parameter estimation that assume time-varying error bounds. A. Set-Membership CCM Stochastic Gradient-Type Algorithm Here we develop the set-membership CCM stochastic gradient-type (SM-CCM-SG) algorithm. The basic idea is to devise a gradient descent strategy to compute a parameter vector wk for user k that minimizes the instantaneous CM cost function subject to certain constraints, which requires for adaptation that the square of the error e2k [i] exceeds a specified error bound γk2 [i]. Mathematically, the proposed SM-CCM-SG algorithm solves the following optimization problem 2 2 2 minimize JCM (wk [i]) = (|wH k [i]r[i]| − 1) = ek [i]

subject to wH k [i]pk [i] = ν whenever

subject to ||hk [i]|| = 1, where R[i] = E[r[i]r H [i]]. The solution is the eigenvector of the Lp × Lp matrix corresponding −1 to the minimum eigenvalue of C H [i]C k obtained by an k R eigenvalue decomposition (EVD). Here, we use Rk [i] in lieu of R[i] to avoid the estimation of both R[i] and Rk [i], which shows no performance loss as reported and investigated in [33], [34], [36]. III. S ET-M EMBERSHIP B LIND A DAPTIVE C ONSTRAINED A LGORITHMS WITH T IME VARYING E RROR B OUNDS In this section, we describe an adaptive filtering framework that combines the set-membership (SM) concept with blind constrained algorithms based on the CCM design. We also introduce simple time-varying error bounds to take into consideration the evolution of the receive filter and the MAI and ISI effects. In SM filtering [2], the parameter vector wk [i] for user k is designed to achieve a specified bound on the magnitude of an estimated quantity zk [i]. As a result of this constraint, the SM blind adaptive algorithm will only perform filter updates for certain data. Let Θ[i] represent the set containing all wk [i] that yield an estimation quantity upper bounded in magnitude by a time-varying error bound γk [i]. Thus, we can write Θ[i] =

∩

{wk ∈ C M :| zk [i] |≤ γk [i]},

(7)

(r[i])∈S

where r[i] is the observation vector, S is the set of all possible data pairs (bk [i], r[i]) and the set Θ[i] is referred to as the feasibility set, and any point in it is a valid estimate zk [i] =

e2k [i]

≥

(8)

γk2 [i]

This problem can be solved using the method of Lagrange multipliers using the equality constraint e2k [i] = γk2 [i] [1]: 2 2 H ∗ L(wk [i], κk ) = (|wH k [i]r[i]| − 1) + [(w k [i]pk [i] − ν)κk ]

+ [κ∗k (pH k [i]w k [i] − ν)], (9) where κk is a Lagrange multiplier. Computing the gradient terms of (9) and equating them to zero, we obtain ∇w∗k [i] L(wk [i], κk ) = 2(ek [i]r[i]zk∗ [i]) + pk [i]κk = 0 ∇κ∗k L(wk [i], κk ) = wH k [i]pk [i] − ν = 0.

(10)

We employ a gradient descent rule to solve for the above equations. Using the first equation of (10), we have wk [i + 1] = wk [i] − µk (ek [i]zk∗ [i]r[i] + pk [i]κk )

(11)

Using the second equation of (10), we obtain −1 wk [i+1] = Πk [i](wk [i]−µk ek [i]zk∗ [i]r[i])+νpk [i](pH , k [i]pk [i]) (12) where µk is the effective step size, ek [i] = |zk [i]|2 −1 is the error −1 H signal for user k, and Πk [i] = I − pk [i](pH pk [i] k [i]pk [i]) is a projection matrix that ensures the constraint and I is an identity matrix. By imposing the condition to update whenever e2k [i] ≥ γk2 [i] we arrive at the set of all wk that satisfy √ √ 1 − γk [i] ≤ |wH 1 + γk [i] (13) k [i + 1]r[i]| ≤

4

It can be verified that the set above is non-convex and comprises two parallel hyper-strips in the parameter space. From the H above conditions we consider two cases: √ √ i) |wk [i + 1]r[i]| ≤ H 1 + γk [i] and ii) |wk [i + 1]r[i]| ≥ 1 − γk [i]. By substituting the recursion obtained in (12) into |wH k [i + 1]r[i]|, we have |wH k [i + 1]r[i]|

µk [i] =

L(wk [i], ϵk ) =

= |zk [i] − µk [i]ek [i]zk [i]r [i]Πk [i]r[i]|. (14) H

Using the above expression we have for case i): √ |zk [i] − µk [i]ek [i]zk [i]r H [i]Πk [i]r[i]| = 1 + γk [i] which leads to (

where λk [i] is a time-varying forgetting factor. Similarly to the case of the SM-CCM-SG algorithm, this problem can be solved with the method of Lagrange multipliers [1] along with the use of the condition e2k [i] ≥ γk2 [i] to save computations

1−

√

1 + γk [i] |zk [i]|

)

(15)

1

(16)

H 2 2 λi−l k [i](|w k [i]r[l]| − 1)

l=1

+

[ϵ∗k (pH k [i]w k [i]

(22)

− ν)],

where ϵk is a Lagrange multiplier. Computing the gradient terms of the Lagrangian in (22) and equating them to zero, we obtain ∇w∗k [i] L(wk [i], ϵk ) =

ek [i]r H [i]Πk [i]r[i]

i ∑

∇ϵ∗k L(wk [i], ϵk ) =

i ∑

2 H λi−l k [i](|zk [l]| − 1)(r[i]r [i]w k [i] + pk [i]ϵk

l=1 wH k [i]pk [i]

− ν = 0. (23)

Using (14) we have for case ii): |zk [i] − µk [i]ek [i]zk [i]r H [i]Πk [i]r[i]| =

√

1 − γk [i]

which results in the following ( ) √ 1 − 1 − γk [i] 1 µk [i] = H |zk [i]| ek [i]r [i]Πk [i]r[i] The resulting SM-CCM-SG algorithm is described by

(17)

(18)

Solving for the above equations, we have [ ( )−1 ˆ k [i] − pH [i]R ˆ −1 [i] d ˆ −1 [i]pk [i] wk [i + 1] = R k k k ) ( ] ˆ k [i] − ν p [i] , ˆ −1 [i]d · pH [i] R k k k

(24)

∑i i−l 2 H ˆ where zk [i] = wH k [i]r[i], Rk [i] = l=1 λk [i]|zk [l]| r[l]r [l], ∑ ˆ k [i] ˆ k [i] = i λi−l [i]z ∗ [l]r[l]. Firstly, we need to compute d d k l=1 k and this is performed by the following recursion

−1 , wk [i+1] = Πk [i](wk [i]−µk [i]ek [i]zk∗ [i]r[i])+νpk [i](pH ˆ k [i] = d ˆ k [i − 1] + λk [i]z ∗ [i]r[i]. k [i]pk [i]) (25) d k (19) where ˆ −1 [i] efficiently and this is At this point, we need to compute R k  ( √ ) done by applying the matrix inversion lemma [1], which yields √  1− 1+γk [i]  1  if |zk [i]| ≥ (1 + γk [i]  |zk [i]|  ek [i]r H [i]Πk [i]r[i]  ˆ −1 [i]r[i]r H [i]R ˆ −1 [i − 1]  ( −1 −1 λk [i]|zk [i]|2 R ) k k ˆ ˆ √ R [i] = R [i−1]− µk [i] = k k √ −1 1− 1−γk [i] 1 2 H ˆ  1 + λk [i]|zk [i]| r [i]Rk [i − 1]r[i] if |zk [i]| ≤ (1 − γk [i]  |zk [i]|  ek [i]r H [i]Πk [i]r[i]  (26)    0 otherwise The last step of the SM-CCM-RLS algorithm is the use of the (20) condition e2k [i] ≥ γk2 [i] to save computations and to adjust the The SM-CCM-SG algorithm described in (19)-(20) requires optimal λk [i]. In order to adjust λk [i], the authors in [6] have M + UR(5M − 1) additions and M + 1 + UR(5M + 4) mul- advocated a strategy that yields bounding ellipsoids that lead to tiplications per received symbol, where UR is the update rate. a simple innovation check with linear complexity, which considers the cost function B. Set-Membership CCM RLS-Type Algorithm ( [ ) ] ek [i] 1 In this part, we derive the set-membership CCM recursive Cλk [i] = λk [i] 2 −1 . γk [i] 1 + λ [i]r H [i]R ˆ −1 [i − 1]r[i] least-squares-type (SM-CCM-RLS) algorithm. The idea is to k k (27) devise a least-squares method to calculate a parameter vector wk used at the receiver for user k that minimizes the The maximization of the cost function in (27) leads to the innoexponentially-weighted CM cost function subject to constraints vation check of the proposed SM-CCM-RLS algorithm:  that require that the squared error of the filter exceed a specified ( )   time-varying error bound γk2 [i]. Mathematically, the proposed |ek [i]| 1 if |e∗k [i]| > γk [i], γk [i] − 1 ˆ −1 [i−1]r[i] r H [i]R λk [i] = SM-CCM-RLS algorithm solves the optimization problem k   0 otherwise. i ∑ (28) i−l LS H 2 2 minimize JCM (wk [i]) = λk [i](|wk [i]r[l]| − 1) The SM-CCM-RLS algorithm described in (24)-(26) and (28) l=1 requires M +UR(4M 2 +3M ) additions and M +2+UR(4M 2 + subject to wH [i]p [i] = ν k k 6M + 3) multiplications per received symbol. It is worth menwhenever e2k [i] ≥ γk2 [i], tioning that the computational savings can be quite substantial (21) if the algorithm operates with a low UR.

5

IV. B LIND PARAMETER E STIMATION AND T IME - VARYING B OUNDS This section presents a blind framework employed to compute time-varying error bounds γk [i] based on parameter and interference dependency. The proposed blind framework is an extension of the approach in [21] that computes time-varying error bounds, and performs interference estimation and tracking. In contrast to [21] that considers training-based recursions, a blind procedure for estimating MAI and ISI power levels is presented with a set-membership blind channel estimator and a blind amplitude estimator, which are employed in the adaptive error bound for the SM adaptive algorithms.

ˆbk [i]

zk [i]

Blind linear receiver wk [i]

Q(·) Aˆk [i] N Blind amplitude estimation

Blind channel estimation ˆ k [i] h RAKE receiver fk [i]

xk [i] dˆ [i] k L-

+

MAI and ISI power estimation

vˆk [i]

Fig. 1. Block diagram of the proposed blind scheme.

A. Parameter Dependent Bound Here, we describe a parameter dependent bound (PDB), that is similar to the one proposed in [13] and considers the evolution of the parameter vector wk [i] for the desired user (user k). The PDB recursion computes a bound for SM adaptive algorithms and is described by √ γk [i + 1] = (1 − β)γk [i] + β α||wk [i]||2 σ ˆ 2 [i],

r[i]

(29)

where β is a forgetting factor that should be adjusted to ensure an appropriate time-averaged estimate of the evolution of the power of the parameter vector wk [i]. The quantity α||wk [i]||2 σ ˆ 2 [i] is the variance of the inner product of wk [i] with n[i] which provides information on the evolution of wk [i], where α is a tuning parameter and σ ˆ 2 [i] is an estimate of the noise power. This kind of recursion helps avoiding too high or low values of the squared norm of wk [i] and provides a smoother evolution of its trajectory for use in the time-varying bound. The noise power at the receiver should be estimated via a time average recursion. In this work, we will assume that it is known at the receiver.

receiver is given by

H xk [i] = f H ck [i] + k [i]r[i] = Ak [i]bk [i]f k [i]˜ | {z } desired signal

K ∑ j=2 j̸=k

|

Aj [i]bj [i]f H cj [i] j [i]˜ {z

}

MAI

+ fH [i]ηk [i] + f H [i]n[i], | k {z } | k {z } noise

ISI

(30) where f H ck [i] = ρk [i] and f H cj [i] = ρ1,j [i] for j ̸= 1. k [i]˜ k [i]˜ The symbol ρk represents the cross-correlation (or inner product) between the effective signature and the RAKE with perfect channel estimates. The symbol ρ1,j [i] represents the crosscorrelation between the RAKE receiver and the effective signature of user j. The second-order statistics of the output of the RAKE in (30) are described by

B. Parameter and Interference Dependent Bound In this part, we develop a blind interference estimation and tracking procedure to be combined with a parameter dependent bound and incorporated into a time-varying error bound for SM recursions. The MAI and ISI power estimation scheme, outlined in Fig. 1, employs both the RAKE receiver and the linear receiver described in (3) for subtracting the desired user signal from r[i] and estimating MAI and ISI power levels. With the aid of adaptive algorithms, we design the linear receiver, estimate the channel modeled as an FIR filter for the RAKE receiver and obtain the detected symbol ˆbk [i], which is combined with an amplitude estimate Aˆk [i] for subtracting the desired signal from the output xk [i] of the RAKE. Then, the difference dk [i] between the desired signal and xk [i] is used to estimate the MAI and ISI power. C. Blind Interference Estimation and Tracking Let us consider the RAKE receiver with perfect channel knowledge, whose parameter vector f k [i] = C k hk [i] for user k (desired one) corresponds to the effective signature sequence ˜k [i] = C k hk [i]. The output of the RAKE at the receiver, i.e. c

E[|xk [i]|2 ] = A2k [i]ρ2k [i] E[|bk [i]|2 ] | {z } →1

+

K K ∑ ∑ j=1 j̸=k

|

l=1 j̸=k

˜j c ˜H A2j [i]E[bj [i]b∗l [i]]f H j c l fj {z

∑ H˜ c H → K j ˜j [i]f j j=1,j̸=k f j s

}

H H H + fH k E[ηk [i]ηk [i]]f k + f k E[n[i]n [i]]f k . | {z } →σ 2 f H k fk

(31) From the previous development, we can identify the sum of the power levels of MAI, ISI and noise terms from the second-order statistics. Our approach is to obtain instantaneous estimates of the MAI, the ISI and the noise from the output of a RAKE receiver, subtract the detected symbol in (3) from this output (using the more reliable linear multiuser receiver (wk [i])) and to track the interference (MAI + ISI + noise) power as shown in Fig. 1. Let us define the difference between the output of the

6

RAKE receiver and the detected symbol for user 1:

where ϕs and ϕn are the signal and noise subspaces, respectively. Since ϕs and ϕn are orthogonal, we have the condition K H H ∑ ϕH n xk [i] = ϕn Ak [i]bk [i]pk [i] = ϕn Ak [i]bk [i]C k hk [i] = 0. dk [i] = xk [i] − Aˆk [i]ˆbk [i] ≈ Ak [i]bk [i]f H sk [i] + f H k [i]˜ k [i]ηk [i] | {z }Hence, we have k=2 ISI | {z } H H ∗ ∗ Γ = hH MAI k [i] C k bk [i]Ak [i]ϕn ϕn Ak [i]bk [i]C k hk [i] | {z } (37) Υ + fH [i]n[i] . k k | {z } noise The above relation allows to blindly estimate the channel hk [i]. (32) To this end, we need to compute the eigenvector corresponding to the smallest eigenvalue of Υk . It turns out that we can use By taking expectations on |dk [i]|2 and taking into account the the fact that limp→∞ (R/σ 2 )−p = ϕn ϕH n [36] and, in practice, assumption that MAI, ISI and noise are uncorrelated we have: it suffices to use p = 1 or 2. Therefore, to blindly estimate the channel of user k in the DS-CDMA system we need to solve the K ∑ H H 2 optimization problem E[|dk [i]| ] ≈ ck [i]˜ ck [i]f k [i] f k [i]˜ k=2 H 2 H + fH k [i]E[ηk [i]ηk [i]]f k [i] + σ f k [i]f k [i], (33)

where the above equation represents the interference power. Based on time averages of the instantaneous values of the interference power, we consider the following algorithm to estimate and track E[|dk [i]|2 ] vˆ[i + 1] = (1 − β)ˆ v [i] + β|dk [i]|2 ,

(34)

where β is a forgetting factor. To incorporate parameter dependency and interference power for computing a more effective bound as an alternative to replace (29), we employ the parameter and interference dependent bound (PIDB) ) (√ √ τ vˆ2 [i] + α||wk ||2 σ ˆ 2 [i] , γk [i + 1] = (1 − β)γk [i] + β (35) where vˆ[i] is the estimated interference power in the multiuser system and τ is a weighting parameter that must be set. The equations in (34) and (35) are time-averaged recursions that √ 2 2 [i] + are aimed at tracking the quantities |d [i]| and ( τ v ˆ k √ α||wk ||2 σ ˆ 2 [i]), respectively. The equations in (34) and (35) also avoid undesirable too high or low instantaneous values which may lead to inappropriate time-varying bound γk [i]. D. Blind Channel and Amplitude Estimation Let us now present a set-membership blind channel estimation (SM-BCE) algorithm to design the RAKE and linear receivers. Consider the constraint matrix C k that contains onechip shifted versions of the signature sequence for user k defined in (2) and the assumption that the symbols bk [i] are independent and identically distributed (i.i.d), and statistically independent from the symbols of the other users. Consider the covariance matrix R = [r[i]r H [i]] and the transmitted signal xk [i] = Ak [i]bk [i]pk [i], where pk [i] = C k hk [i]. Let us now perform an eigen-decomposition on R R=

K ∑

H 2 E[xk [i]xH k [i]] + E[η k [i]η k [i]] + σ I

k=1

[ ] = ϕs ϕn

[

Λs + σ 2 I 0

0 σ2 I

]

[ ]H ϕs ϕn ,

(36)

ˆ k [i] = arg min hH [i]Υk hk [i], subject to ||hk [i]|| = 1, h k hk [i]

(38) In order to solve (38) efficiently, we rely on the SM estimation strategy and a variant of the power method [33] that uses a simple shift is adopted to yield the SM-BCE ˆ k [i] = (I − τk [i]Υ ˆ k [i − 1], ˆ k [i])h h

(39)

ˆ k [i] ← h ˆ k [i]/||h ˆ k [i]|| to norˆ k [i]] and h where τk [i] = 1/tr[Υ ˆ malize the channel. The quantity Υk [i] is estimated by ˆ p [i]Aˆk [i]C k , ˆ k [i] = Υ ˆ k [i − 1] + λk [i]C H Aˆ∗ [i]P Υ k k k

(40)

where λk [i] is a variable forgetting factor that is obtained by (28) ˆ p [i] is computed according to (26). Next, we describe a and P k procedure to estimate the amplitude. In general, amplitude estimation is an important task at the receiver that is useful for interference cancellation and power control. The proposed blind interference estimation and tracking algorithm needs some form of amplitude estimation in order to accurately compute the interference power. To estimate the amplitudes of the associated user signals, we describe the following procedure to estimate the absolute value of the output of the RAKE receiver defined in (30) as given by qk [i + 1] = (1 − β)qk [i] + |xk [i]|.

(41)

The amplitude can be estimated by removing the square-root of the interference power from the above estimate according to Aˆk [i + 1] = (1 − β)Aˆk [i] + (|qk [i]| − |vk [i]|)

(42)

The above procedure is simple and effective to estimate the amplitude for use in the interference power estimation procedure. Since one recursion depends on the other, a designer shall start the procedure with an interference power equal to zero (or equivalently |vk [i]| = 0). V. A NALYSIS OF T HE A LGORITHMS In this section, we study of the properties of the optimization problems associated with the design of the SM-CCM-based algorithms and examine the convergence and tracking performances of the proposed algorithms. In order to ensure the convergence of the SM-CCM-RLS a persistence of excitation condition on the received data r[i] must hold and the transmitted

7

symbols bk [i] and the noise n[i] have to be uncorrelated. To this end, the variable bound must adapt such that the above conditions are met, following the same procedure established for OBE algorithms in [18]. Since the SM-CCM-RLS algorithm is expected to converge without misadjustment to the Wiener filter and certain aspects of the convergence analysis of the SMCCM-SG algorithm are of greater interest, we will focus on the latter. The CCM-based algorithms are inherently nonlinear and deal with time-varying environments, which leads to difficulties in the study of their performance. For this reason, we will resort to an effective approach termed energy conservation principle [38] that lends itself to such analysis.

2 2 minimize JCM (wk [i]) = E[(wH k [i]r[i]| − 1) ]

= E[|zk [i]|4 − 2|zk [i]|2 + 1] subject to wH k [i]pk [i] = ν 2 2 2 whenever (|wH k [i]pk [i]| − 1) ≥ γk [i]

(43) Let us rewrite the received vector as

k=1

Ak bk [i]pk [i] +η k [i] + n[i], | {z }

(44)

xk [i]

where pk [i] = C k hk [i], R = S + G + σ 2 I, S = ∑K ∑K H E[ k=1 xk [i]xH k [i]], and G = E[ k=1 η k [i]η k [i]], bk [i] are independent and identically distributed random variables, and are statistically independent from n[i]. Consider user k as the desired user and let wk [i] be the receive filter for this user. Our strategy to analyze the above optimization problem and its properties is to transform the variables and rewrite the problem in a convenient form that provides more insight about the nature of the problem. Let us now define the signal of the desired and the signal of all the users after applying the receive filter wk [i]: H T tk , Ak pH k w k [i], and t , AP w k = [t1 . . . tK ] , (45)

where P , [p1 . . . pK ], A , diag[A1 . . . AK ] and b , [b1 . . . bK ]T . The relation between the receive filter, the channel and the signature sequence can be written as H wH k [i]pk [i] = w k [i]C k hk [i] = ν.

= E[(tH bbH t)2 ] − 2E[(tH bbH t] + 1 K K K ∑ ∑ ∑ = 8( tj t∗j )2 − 4 (tj t∗j )2 − 4 tj t∗j + 1 j=1

= 8(D +

j=1 K ∑

tj t∗j )2

j=1

− 4D − 4 2

j=1

K ∑

(48)

(tj t∗j )2

j=2 K ∑

(tj t∗j ) + 1,

j=1,j̸=k

Let us consider the optimization problem that needs to be solved for the design of the blind receiver and can be solved by the algorithms proposed in Section III:

K ∑

JCM (t) = E[|zk [i]|4 − 2|zk [i]|2 + 1]

− 4D − 4

A. Analysis of the Optimization Problem

r[i] =

By considering the noise and the ISI negligible we can write the cost function in (43) as

(46)

Therefore, we have for the desired user k the following equivalences H H H tk [i] = Ak pH k [i]w k [i] = Ak hk [i] C k w k [i] = Ak hk [i]hk [i]. | {z }

where t = f (wk [i]) is a linear function of the receive filter wk [i], the terms multiplying the summations in the third line of (48) are obtained by evaluating the expected values in the second ˆ H [i]hk [i]|. The line of (48) [31], and D = tj t∗j = ν 2 |Ak |2 |h k strategy we employ to enforce the convexity of the optimization problem relies on the adjustment of the parameter ν and a transformation of variables on the cost function in (48), which will be detailed in what follows. Let us now transform the above cost function taking into account the constraint wH k [i]pk [i] = ν and the fact that we are interested in demodulating user k and rejecting the remaining users. Therefore, we introduce another parameter ¯ k [i] that exvector ¯t = [t1 . . . tk−1 tk+1 . . . tK ]T = Bw cludes the user k and is responsible for the interference suppression task of the remaining users (all but user k), where ¯ = A ¯P ¯ H, P ¯ [i] = [p1 [i] . . . pk−1 [i] pk+1 [i] . . . pK [i]] and B A = diag[A1 . . . Ak−1 Ak+1 . . . AK ]. The transformed cost function is given by H H JCM (¯t) = 8(D + ¯t ¯t)2 − 4(D2 + ¯t ¯t)2 H − 4(D + ¯t ¯t) + 1.

(49)

At this point we need to take into account the constraint ¯ k [i] is a linear mapping JCM (wk [i]) ≤ γk2 [i]. Since ¯t = Bw we have an equivalent constraint JCM (¯t) ≤ γ¯k2 [i], where γ¯k2 [i] is the bound modified by the linear mapping. It can be verified that the constraint set generated by JCM (¯t) ≤ γ¯k2 [i] is not convex and leads to two disjoint parallel hyperstrips in the parameter space. Given that the constraint is not convex, the optimization problem is clearly non convex. However, in this context nonconvexity only poses a problem if local minima are present and prevent an algorithm from reaching the global minimum. It turns out that a designer can adjust the parameter ν in order to enforce the convexity for each hyperstrip. Computing the Hes(t) we obtain sian [43] H = ∂¯t∂H ∂JCM ∂¯ t [ H = 16 (D − 1/4)I+

( )] ¯tH ¯tI + ¯t¯tH − diag |t1 [i]|2 . . . |tk−1 |2 |tk+1 |2 . . . |tK [i]|2 . | {z } positive or positive semi−definite terms

hk [i]

(47)

(50)

8

ˆ H [i]h ˆ k [i]|2 ≥ 1/4 ensures that The condition D = ν 2 |Ak |2 |h k there is no local minimum in each of the hyperstrips resulting from the optimization problem and convexity can be enforced ¯ k [i] is a linear funcin each of the hyperstrips. Since t = Bw tion of wk [i] then JCM (¯t) preserves the maxima and minima properties of JCM (wk [i]). For sufficiently high signal-to-noise ratio (SNR) values, the extrema of the cost function can be considered a small perturbation of the noise-free case [31]. B. Stability Analysis In this part, we discuss the stability analysis of the SM-CCMSG algorithm described in subsection III.A. In particular, we consider the range of step-size values for convergence. Let us now rewrite the update equation of the algorithm as

stability, a sufficient condition for (55) to hold implies that 0 ≤ E[µ(∞)] < min k

2 , |λvr k |

(56)

where λvr k is the kth eigenvalue of Rvr that is not necessarily real since Rrv is not symmetric. C. Steady-State Analysis In this part of the analysis we are interested in devising a formula to predict the excess MSE, which depends on the MAI, the ISI, the noise at the receiver and the parameters of the SMCCM-SG algorithm. The excess MSE is related to the error in the filter coefficients ϵwk [i] via the a priori estimation error, which is defined as

−1 ea [i] , ϵH (57) wk [i + 1] = Πk [i](wk [i] − µk [i]ek [i]zk∗ [i]r[i]) + νpk [i](pH wk [i]r[i], k [i]pk [i]) H ( pk [i]pk [i] ) = wk [i] − µk [i]ek [i]r H [i]wk [i] I − H r[i] where ϵwk [i] = wopt − wk [i], and wopt is the optimum linear pk [i]pk [i] MMSE receiver. Consider the MSE at time i: H = wk [i] − µk [i]ek [i]v[i]r [i]wk [i] 2 MSE[i] = E[|bk [i] − wH k [i]r[i]| ] H = (I − µk [i]ek [i]v[i]r [i])wk [i], H = ϵmin + E[|ea [i]|2 ] + pH k [i]E[ϵwk [i]] + E[ϵwk [i]]pk [i] (51) H H H − E[wH opt r[i]r [i]ϵwk [i]] − E[ϵwk [i]r[i]r [i]w k,opt ], H p [i]p [i] (58) where v [i] = (I − k k r[i]) and the expression in the k

pH k [i]pk [i]

second line of (51) is obtained by substituting Πk [i] = I − −1 H pk [i](pH pk [i] into the first line and further manipuk [i]pk [i]) lating the terms. Let us now define the error vector

H where ea [i] , ϵH wk [i]r[i] = (w opt − w k [i]) r[i] and ϵmin = H 2 E[|bk [i] − wopt r[i]| ]. When i → ∞, we have wk [i] → wk,opt and E[ϵwk [i]] → 0 and the steady-state MSE

ϵwk [i + 1] = wk,opt − wk [i + 1]

lim MSE[i] = ϵmin + lim E[|ea [i]|2 ]

i→∞

i→∞

(59)

= (I − µk [i]ek [i]v k [i]r [i])ϵwk [i] + µk [i]ek [i]v k [i]r [i]wk,opt . (52) The steady-state excess MSE is then defined [1] as H

H

In order to proceed with the analysis, we need to resort to an assumption. Assumption 1: Let us suppose that for the algorithm in (51) when i → ∞ E[µk [i]e2k [i]] = E[µk [i]]E[e2k [i]].

(53)

This assumption holds if µk [i] is a constant, and we claim that it is approximately true if µk [i] varies slowly around its mean value. By writing E[µk [i]e2k [i]] = E[µk [i]]E[e2k [i]] + E[(µk [i] − E[µk [i]])e2k [i]], (54) we can notice that the second term on the right-hand side will be small compared with the first one provided that µk [i] varies slowly around its mean value. By taking expectations on both sides of (52) and using the previous assumption, we have E[ϵwk [i + 1]] = (I − E[µk [i]]Rrv )E[ϵwk [i]],

(55)

where Rrv = E[ek [i]v k [i]r H [i]] and Rvr wopt = 0. From the above, it can be concluded that w [i] con∏k∞ verges to wk,opt and (55) is stable if and only if i=0 (I − E[µk [i]]Rvr ) → 0, which is a necessary and sufficient condition for limi→∞ E[ϵwk ] = 0 and E[wk [i]] → wk,opt . For

ξ , lim E[|ea [i]|2 ]

(60)

i→∞

Using the energy conservation principle [38], the proposed SMCCM-SG algorithm can be written in the form ( ) pk [i]pH ∗ k [i] wk [i + 1] = wk [i] + µk [i] (−ek [i]zk [i]) I − H r[i], {z } | pk [i]pk [i] | {z } Fe [i] k

uk [i]

(61) where Fek [i] is a generic scalar function determined by the adaptive algorithm. Subtracting the above recursion for wk [i + 1] from wk,opt , we obtain ϵwk [i + 1] = ϵwk [i] − µk [i]uk [i]Fek [i]

(62)

Using the a priori estimation error ea [i] , ϵH wk [i]r[i]. Rewriting the previous equation, we obtain ∗ ea [i] = ep [i] + µk [i]uH k [i]r[i]Fek [i] ( ) (63) pk [i]pH H k [i] = ep [i] + µk [i]r [i] I − H r[i]Fe∗k [i] pk [i]pk [i] ( ) p [i]pH [i] Since r H [i] I − pkH [i]pk [i] r[i] = uH k [i]uk [i], we have k

k

∗ ep [i] = ea [i] − µk [i]uH k [i]uk [i]Fek [i].

(64)

9

Using the fact that Fe∗k [i] = ϵwk [i] as ϵwk [i + 1] = ϵwk [i] −

ea [i]−ep [i] , µk [i]uH k [i]uk [i]

uk [i] [e∗ [i] H uk [i]uk [i] a

we can rewrite

− e∗p [i]]

(65)

Rearranging, we obtain ϵwk [i+1]+

uk [i] uk [i] e∗p [i] = ϵwk [i]+ H e∗a [i] (66) uH [i]u [i] u [i]u [i] k k k k

By defining µ ¯[i] = 1/(uH k [i]uk [i]), squaring the previous equation and taking expectations, we obtain the relation

2 where A = 3 + 3σ 4 MAI + 6σMAI σv2 + 6σMAI σn2 + 3σv4 + 2 2 4 6 2 4 2 6σv σn + 3σn and B = ση + 3σv ση + 3σMAI ση4 + σv6 + 6σv2 + 2 2 4 2 4 2 6 4 ση σMAI + 3σMAI σn + 3σMAI σv + σMAI + σMAI + 2σv2 σn2 + 2 2 4 2 σv4 + 2σMAI + 2σMAI σv2 + σMAI + 4σv2 + 2ση2 + 2σMAI + 2. Upon convergence(i → ∞), we can assume that 2m E[MAI2m ](E[MAI2 ])m = σMAI , E[η 2m ] = (E[η 2 ])m = ση2m , and E[v 2m ] = (E[v 2 ])m = σv2m , where σMAI , ση , and σv are the variances of the Gaussian distribution. In this situation, the high power terms E[|ea [i]|4 ] and E[|ea [i]|6 ] may be neglected as these values are typically very small compared with the remaining terms. The excess MSE is obtained as follows

2 E[µ2k (∞)]E[uH k [i]uk [i]]B E[||ϵwk [i+1]||2 ]+E[¯ µk [i]|ep [i]|2 ] = E[||ϵwk [i+1]||2 ]+E[¯ µk [i]|eaξ[i]| ], a [i]|2 ] = = E[|e 2 2E[µk (∞)](σMAI + σv2 + ση2 ) − E[µ2k (∞)]E[uH (67) k [i]u (72) In the steady state, we can write

E[||ϵwk [i]||2 ] = E[||ϵwk [i+1]||2 ], and E[¯ µk [i]|ea [i]|2 ] = E[¯ µk [i]|e [i]|2 ] D. pTracking Analysis (68) In this subsection, we assess the proposed SM-CCM-SG alThe energy preserving equation can be used to compute gorithm in a non-stationary environment, in which the algorithm the excess MSE. It can be obtained by cancelling the terms has to track the minimum point of the error-performance surE[||ϵwk [i + 1]||2 ] and E[||ϵwk [i]||2 ] in (67) and by substituting face. Specifically, we derive an expression for the excess MSE (64) into (67), which yields of a blind adaptive linear receiver when the channel varies in E[¯ µk [i]|ea [i]|2 ] = E[¯ µk [i]|ea [i] −

µk [i] ∗ 2 F [i]| ], µ ¯k [i] ek

(69)

where Fek [i] = −ek [i]zk∗ [i] = (1 − |zk [i]|2 )zk∗ [i] and zk [i] = (wk,opt −ϵwk [i])H r[i] = wH k,opt r[i]−ea [i] = Ak bk [i]+MAI+ η k [i] + v[i] − ea [i]. Substituting Fek [i] and (64) into (69) and manipulating the terms, we obtain E[µk [i]]E[ea [i]zk [i](1 − |zk [i]|2 )] +E[µk [i]]E[µk [i]]E[ea [i]zk∗ [i](1 − |zk [i]|2 )] {z } | =

C 2 E[µ2k [i]]E[uH [i]u k [i]|zk [i]| (1 k

|

{z

(70)

− |zk [i]| ) ] . } 2 2

F

At this point we need to resort to another assumption to continue with our analysis. 2 Assumption 2: In the steady state, uH k [i]uk [i] and |Fek [i]| are uncorrelated. The quantities {bk [i], MAI, η[i], v[i], ea [i]} are zero mean random variables, and are mutually independent. We use the fact that E[b2m k ] = 1 for any positive integer m and that the residual MAI and ISI are Gaussian random variables. Using the previous assumption and substituting zk [i] = wH k [i]r[i] into (70), we obtain 2 E[µ2k [i]]E[uH k [i]uk [i]]AE[|ea [i]| ] 2 4 + 3E[µ2k [i]]E[uH k [i]uk [i]]E[MAI ]E[|ea [i]| ] 2 4 + 3E[µ2k [i]]E[uH k [i]uk [i]]E[v [i]]E[|ea [i]| ]

time. Differently from the works in [38], [40], where expressions were derived for the constant modulus (CM) algorithm and the CCM algorithm, respectively, we consider a set-membership approach. In the time-varying scenarios of interest, the optimum receive filter coefficients are assumed to vary according to the model wopt [i + 1] = wopt + q[i], where q[i] denotes a random perturbation. This is consistent with tracking analyses of adaptive filtering algorithms and requires an assumption. Assumption 3: The sequence q[i] is a stationary sequence of independent zero-mean vectors and positive definite autocorrelation matrix Q = E[q[i]q H [i]], which is mutually independent of the sequences {uk [i]}, {v[i]}, {MAI[i]}, {η[i]}. Let us consider the weight-error vector ϵwk [i] = wk,opt − wk [i], which satisfies ϵwk [i + 1] = ϵwk [i] − µk [i]uk [i]Fek [i] + q[i].

(73)

H With ea [i] = ϵH wk [i]r[i] and ep [i] = ϵwk [i + 1]r[i], we have

ea [i] = ep [i] + µk [i]||uk [i]||2 Fe∗k [i].

(74)

Using (73) and (74), we obtain ϵwk [i + 1] + µ ¯k [i]uk [i]e∗a [i] = ϵwk [i] + q[i] + µ ¯k [i]uk [i]e∗p [i]. (75) Squaring (75) and taking the expected value on both sides, we obtain E[||ϵwk [i + 1]||2 ] + E[¯ µk [i]|ea [i]|2 ] = E[||ϵwk [i] + q[i]||2 ]

2 4 + E[µ2 [i]]E[uH k [i]uk [i]](3E[µ [i]] + 1)E[|ea [i]| ]

+ E[¯ µk [i]|ep [i]|2 ], (76)

2 H 6 E[µ2k [i]]E[uH k [i]uk [i]]B + E[µk [i]]E[uk [i]uk [i]]E[|ea [i]| ] {z } | C

where

2

= 2E[µk [i]](E[MAI ]E[|ea [i]|2 ] + E[v 2 [i]]E[|ea [i]|2 ]

E[||ϵwk [i] + q[i]||2 ] = E[||ϵwk [i]||2 ] + E[ϵwk [i] + q[i]]

+E[η 2 [i]]E[|ea [i]|2 ] + E[|ea [i]|4 ], | {z } D

(71)

+ E[q H [i]ϵwk [i]] + E[q H [i]q[i]]. (77)

10

E[¯ µk [i]|ea [i]|2 ] = Tr(Q) + E[¯ µk [i]|ea [i] −

µk [i] ∗ 2 F [i]| (78) µ ¯k [i] ek

Expanding the equation above, it can be simplified to C = Tr(Q) + D,

(79)

where C and D are defined in (71). The excess MSE is then obtained as ξ = E[|ea [i]|2 ] =

E[µ2k (∞)]E[||uk [i]||2 ]B + Tr(Q) 2 + σv2 + ση2 ) − E[µ2k (∞)]E[|uk [i]|2 ]A 2E[µk (∞)(σMAI (80)

E. Computation of Moments In order to compute the expressions obtained in (72) and (80), we need to obtain the first and second-order moments of µk [∞]. To this end, we resort to the expression for the variable step size in (20) and the methodology employed in [3], which after some algebraic manipulations yields E[µk [∞]] = E[γk [i]]Pup +

(1 − Pup ) E[Ak [i]] , E[γk [i]] E[|uk [i]|2 ]

(81)

(1 − Pup ) E[A2k [i]] , (82) E[γk [i]] E[|uk [i]|4 ] where the probability of update Pup is given by [ ] [ ] Pup = P r |ek [i]| > E[γk [i]] = P r |ek [i]|2 > E[|γk [i]|2 ) ( E[γk [i]] , = 2Q σe (83) E[µ2k [∞]] = E[γk [i]]Pup +

where P r[·] denotes the probability Q(x) is the complementary Gaussian cumulative distribution function [44] given by ∫ ∞ 2 1 √ e−t /2 dt. Q(x) = (84) 2π x

by hk,l [i] = pk,l αk,l [i], where αk,l [i] (l = 0, 1, . . . , Lp − 1, k = 1, 2, . . . , K) are obtained with Clarke’s model [44] and pk,l represent the powers of each channel path. We show the results in terms of the normalized Doppler frequency fd T (cycles/symbol), where fd is the maximum Doppler shift and T is the symbol interval. We use three-path channels with relative powers given by 0, −3 and −6 dB,where in each run the spacing between paths is obtained from a discrete uniform random variable between 1 and 2 chips. The proposed blind channel estimator described in Section IV and that of [41] model the channel as a finite impulse response (FIR) filter and we employ a filter with 6 taps as an upper bound for the experiments. The phase ambiguity derived from the blind channel estimation method in [41] ˆ k (0) as is addressed in our simulations by using the phase of h a reference to remove the ambiguity and for time-varying channels we assume ideal phase tracking. Alternatively, differential modulation can be used to account for the phase rotations. The tuning parameter α, the forgetting factor β and the weighting parameter τ required by the time-varying bounds described in Section IV have been obtained by experimentation and chosen such that the performance of the algorithms is optimized. The update rate (UR) has been computed by counting, for each simulation trial t, the number of updates (Nu,t ) performed and then dividing it by the number of received symbols (Ns,t ). Then, the ∑T N , where T = 1000 is the total U R is given by U R = T1 t Nu,t s,t number of trials. A. MSE Analytical Results The aim of this part is to verify the validity of the analytical results obtained in Section V. Specifically, we shall evaluate the analytical MSE obtained with the analytical formulas in (72) and (80), and compare them to the results obtained by simulations. We consider first a scenario with fixed channels (fd = 0) and assess the MSE curves using (72) against the number of received symbols and also versus the signal-to-noise ratio (SNR) defined by Eb /N0 , as shown in Fig. 2. The results show that the analytical curves match very well those obtained via simulations, showing the validity of our analysis and assumptions. a) N=31, K=10 users, Eb/N0 =15 dB

VI. S IMULATION R ESULTS

−5

Simulated(α=12, τ=0.4, β=0.98) Analytical Simulated(α=6, τ=0.35, β=0.98) Analytical Optimal

0 −2

−10

−4 MSE (dB)

In this section we assess the performance of the proposed and analyzed adaptive algorithms in terms of mean-square error (MSE) and bit error rate (BER). In particular, we consider the proposed SM-CCM-SG and SM-CCM-RLS adaptive algorithms and the existing CMV-SG and CMV-RLS reported in [26], the CCM-SG [32] and the CCM-RLS [33] algorithms, the SM-CMV-SG and SM-CMV-RLS reported in [7] with and without the proposed blind PDB and PIDB time-varying bounds described in Section IV. The DS-CDMA network employs Gold sequences of length N = 31, the users are randomly distributed and communicate over multipath channels. The channels experienced by different users are independent and different since we focus on an uplink scenario. The DS-CDMA system under consideration employs quadrature phase-shift keying (QPSK) modulation. The channel coefficients are given

b) N=31, K=10 users

2

MSE (dB)

H Using Assumption 3, we have E[ϵH wk [i]q[i]] = E[q [i]ϵwk [i] = 2 0. When i → ∞ E[||ϵwk [i + 1]|| ] = E[||ϵwk [i]||2 ], the energy preserving equation describing the tracking performance is given by

−6

−15

−8 −10

−20

−12 −14

0

50 100 150 200 Number of received symbols

250

−25 10

15 20 Eb/N0 (dB)

25

Fig. 2. MSE performance against a) the number of received symbols at Eb /N0 = 15dB and b) Eb /N0 (dB).

11

The tracking analysis of the proposed SM-CCM-SG algorithm in a time-varying multipath fading channel is discussed next. We consider the same scenario as before except for the fact that the channel is now time-varying and the analytical results of the tracking analysis in (80) are employed. We consider a Jakes’ model with a typical normalized fading rate fd T . We compute Tr(Q) with the aid of J0 (2πfd T ) [44], which is the autocorrelation function of the Jakes’ model and where J0 is the zero-order Bessel function of the first kind. A comparison of the curves obtained by simulations and by the analytical formulas is shown in Fig. 3. Similarly to the case of fixed channels, the analytical and simulated curves agree and show the validity of the proposed formulas. Nevertheless, the curves indicate a higher misadjustment as compared to the curves in Fig. 2 due to the time-varying process and extra effort of the adaptive filters to track the channel variations. a) N=31, K=10 users, Eb/N0=15 dB, fdT=10−4 2

b) N=31, K=10 users −5

Simulated(α=5,τ=0.25, β=0.98) Analytical Optimal

0 −2

−10

MSE (dB)

MSE (dB)

−4 −6

−15

−8 −10

−20

k

−12 −14

0

50 100 150 200 Number of received symbols

−25 10

250

At this point, we wish to evaluate the effectiveness of the proposed algorithms for estimating and tracking the interference power and the amplitude. To this end, we have carried out an experiment, depicted in Fig. 4, where the proposed algorithm estimates of the MAI and ISI powers and the amplitude of the desired user are compared to the actual interference power and amplitude. The results show that the proposed algorithms are very effective for estimating and tracking the interference power in dynamic environments, as depicted in Fig. 4. We assess the proposed channel estimation (CE) algorithm, called SM-CCM-CE, with the time-varying bounds PDB and PIDB, and also with a fixed bound, and compare them to the CMV-based method (CMV-CE) reported in [41] and the subspace algorithm of [42] in terms of the MSE between the actual and the estimated channels using the following dynamic scenario. The system has initially 10 users, the power distribution among the interferers follows a log-normal distribution with associated standard deviation of 3 dB. After 1500 symbols, 6 additional users enter the system and the power distribution among interferers is loosen to 6 dB. The results, shown in Fig 5, reveal that the proposed SM-CCM-CE algorithm outperforms the CMV-CE technique reported in [41] because SMCCM-CE uses a variable forgetting factor. The proposed SMCCM is also computationally simpler than CMV-CE due to the sparse updates, whereas it is substantially less complex than the subspace method of [42]. This is because the SM-CCM-CE needs to compute the principal eigenvector of the Lp × Lp maˆ −1 [i]C k , while the subspace technique in [42] requires trix C k R

15 20 Eb/N0 (dB)

25

ˆ −1 [i]. Since the principal eigenvector of the M × M matrix R Lp ≪ M in most practical scenarios SM-CCM-CE is typically considerably simpler than the subspace method. N=31, Eb/N0 = 15 dB, fdT = 0.0001

Fig. 3. MSE performance against a) the number of received symbols at Eb /N0 = 15dB and b) Eb /N0 (dB).

SM−CCM−CE−UR=20% SM−CCM−CE−PDB−UR−17.8% SM−CCM−CE−PIDB−UR−16.7% CMV−CE Subspace

−1

10

a) N=31, Eb/N0 = 12 dB, K=10, fdT=5x10−3 b) N=31, E /N = 12 dB, K=10, f T=5x10−3 b 0 d 0.2

2 Interference Power Interference Power Estimate

0.18

1.5

0.16 0.14

1 Amplitude value

Interference Power

MSE (dB)

B. Interference Power and Channel Estimation

0.12 0.1 0.08 0.06 0.04

−2

10

0

500

1000 1500 2000 Number of received symbols

2500

3000

0.5

Fig. 5. MSE performance of channel estimation versus number of received symbols in a dynamic scenario where receivers operate at SN R = Eb /N0 = 15 dB for the desired user.

0

−0.5

0.02 0

0

500 1000 1500 Number of received symbols

−1

C. BER Performance 0

500 1000 1500 Number of received symbols

Fig. 4. Performance of the a) interference power estimation and tracking and b) amplitude estimation algorithms at Eb /N0 = 12dB.

The SM-CCM algorithms are assessed in a non-stationary environment where users enter and exit the system, as depicted in Figs. 6 and 7. The system starts with 2 interferers with 7 dB above the desired user’s power level and 5 interferers with the

12

b

0

d

SM−CCM−RLS−UR=12.4% SM−CMV−RLS−UR=12.1% CCM−RLS CMV−RLS

−2

10

−3

10

0

500

1000 1500 2000 Number of received symbols

2500

3000

Fig. 7. BER performance of the algorithms versus number of symbols for a non-stationary scenario. The parameters are γ = 1.3 for the SM-CMV, γ = 0.65 for the SM-CCM.

sults in the smallest number of updates, followed by the PDB approach and the fixed bound. N=31, Eb/N0=15 dB, fdT=0.0001 SM−CCM−SG−UR=20.4% SM−CCM−SG−PDB−UR=14.2% SM−CCM−SG−PIDB−UR=8.3%

−1

10

N=31, Eb/N0=15 dB, fdT=0.0001

−1

N=31, E /N = 15 dB, f T=0.0001

−1

10

BER

same power level of the desired one, which corresponds to the signal-to-noise ratio Eb /N0 = 15 dB. At 1000 symbols, 2 interferers with 10 dB above the desired signal power level and 1 with the same power level enter the system, whereas 1 interferer with 7 dB above the desired signal power level leaves it. At 2000 symbols, 1 interferer with 10 dB above, and 5 interferers with the same power level of the desired signal exit the system, while 1 interferer with 15 dB above the desired user enters the system. The results for 100 runs show that the proposed SM-CCM-RLS algorithm achieves the best performance, followed by the proposed SM-CMV-RLS recursion, the CCM-SG and the CMV-SG methods. In summary the SM-CCM algorithms outperform the SM-CMV techniques in all scenarios and the SM-CCM-RLS algorithm is the best among the analyzed algorithms. In a near-far scenario the eigenvalue spread of the covariance matrix of the received vector r[i] is large, affecting the convergence performance of the SG algorithms with fixed step size and making it very difficult to compute a pre-determined value [1] for the step size. In this case, the SM-CCM algorithms are able to deal with near-far situations since they adopt variable step size or variable forgetting factor mechanisms, ensuring good tracking performance in dynamic scenarios and an improved performance over the existing CCM-type algorithms. In addition, due to their data-selective update feature the SM algorithms can save significant computational resources as they only require parameter updates for about 20% of the time for the SG-type recursions and around 12% for the RLS-based techniques.

BER

10

−2

BER

10 −2

10

−3

SM−CCM−SG−UR=20.4% SM−CMV−SG−UR=19.1% CCM−SG CMV−SG

−3

10

0

500

1000 1500 2000 Number of received symbols

2500

3000

Fig. 6. BER performance of the algorithms versus number of symbols for a non-stationary scenario. The parameters are γ = 1.3 for the SM-CMV, γ = 0.65 for the SM-CCM.

The same scenario illustrated in Fig. 6 is considered for the SM-CCM algorithms with time-varying error bounds, as shown in Figs. 8 and 9. The results indicate that the proposed timevarying bounds are capable of improving the performance of SM algorithms, while further reducing the number of updates. The RLS-type algorithms outperform the SG-based techniques as expected and have lower U R. Moreover, the algorithms with the PIDB approach achieve the best performance, followed by the algorithms with the PDB method and the SM recursions with fixed bounds. With respect to the U R, the PIDB technique re-

10

0

500

1000 1500 2000 Number of received symbols

2500

3000

Fig. 8. BER performance of the SG algorithms versus number of symbols for a non-stationary scenario with time-varying bounds. The parameters are α = 8, τ = 0.35 and β = 0.95 for the time-varying bounds.

The BER performance versus Eb /N0 and number of users is shown in Fig. 10. We consider data packets of P = 1500 symbols and compare the proposed SM-CCM-RLS algorithm with a fixed bound, with the PIDB mechanism, the training-based BEACON algorithm with the PIDB mechanism [21] and the linear MMSE receiver that assumes perfect knowledge of the channels and the noise variance. We have measured the BER after 200 independent transmissions and the BEACON algorithm [6],[21] is trained with 200 symbols and is then switched to decision-directed mode. The curves illustrate that the proposed blind SM-CCM-RLS algorithm can approach the performance of the supervised BEACON algorithm and that of the linear MMSE receiver.

13

N=31, E /N = 15 dB, f T=0.0001 b

−1

0

ACKNOWLEDGEMENT

d

10

SM−CCM−RLS−UR=12.4% SM−CCM−RLS−PDB−UR=10.2% SM−CCM−RLS−PIDB−UR=5.4%

The second author is grateful for the financial support provided by CNPq, a Brazilian research council, and FAPERJ, a Rio de Janeiro state research funding agency.

BER

R EFERENCES [1]

−2

10

[2] [3] [4] −3

10

0

500

1000 1500 2000 Number of received symbols

2500

3000

Fig. 9. BER performance of the RLS algorithms versus number of symbols for a non-stationary scenario with time-varying bounds. The parameters are α = 8, τ = 0.35 and β = 0.95 for the time-varying bounds.

[5]

[6] [7]

N=31, K=10 users, fdT=0.0001

−1

N=31, Eb/N0 = 15 dB, fdT=0.0001

−1

10

10

[8] [9] −2

−2

10

[10]

BER

BER

10

[11] −3

−3

10

10 SM−CCM−RLS−UR−20.4% SM−CCM−RLS−PIDB−UR−15.4% BEACON−PIDB−UR−15.6% MMSE 5

10 15 Eb/N0 (dB)

20

[12]

2

4

6 8 10 12 14 16 18 20 Number of users (K)

Fig. 10. BER performance versus (a) Eb /N0 with K = 10 users and (b) number of users (K) at Eb /N0 = 15 dB .

[13] [14] [15] [16]

VII. C ONCLUSIONS We have proposed SM-CCM adaptive algorithms for interference suppression in DS-CDMA systems. We have analyzed the optimization problem that gives rise to the SM-CCM algorithms and devised analytical expressions to predict their MSE performance with a good accuracy. The proposed SM-CCM algorithms are equipped with variable step sizes and forgetting factors and are only updated for a small fraction of time without incurring any performance degradation. A blind framework for SM estimation that takes into account parameter estimation dependency and MAI and ISI for multiuser communications has also been introduced to provide a time-varying bound that is robust against channel and SNR variations. Simulations have shown that the proposed SM-CCM algorithms outperform previously reported blind techniques for several scenarios.

[17]

[18]

[19] [20]

[21]

P. S. R. Diniz, Adaptive Filtering: Algorithms and Practical Implementations, 2nd edition, Kluwer Academic Publishers, Boston, MA, 2002. S. Gollamudi, S. Nagaraj, S. Kapoor and Y. F. Huang, “Set-Membership Filtering and a Set-Membership Normalized LMS Algorithm with an Adaptive Step Size”, IEEE Sig. Proc. Letters, vol. 5, No. 5, May 1998. S. Werner and P. S. R. Diniz, “Set-membership affine projection algorithm”, IEEE Sig. Proc. Letters, vol 8., no. 8, August 2001. P. S. R. Diniz and S. Werner, “Set-Membership Binormalized Data Reusing LMS Algorithms,” IEEE Trans. Sig. Proc., vol. 51, no. 1, pp. 124134, January 2003. S. Gollamudi, S. Kapoor, S. Nagaraj and Y. F. Huang, “Set-Membership Adaptive Equalization and an Updator-Shared Implementation for Multiple Channel Communication Systems”, IEEE Transactions on Signal Processing, vol. 46, no. 9, pp. 2372-2385, September 1998. S. Nagaraj, S. Gollamudi, S. Kapoor and Y. F. Huang, “BEACON: an adaptive set-membership filtering technique with sparse updates”, IEEE Transactions on Signal Processing, vol. 47, no. 11, November 1999. S. Nagaraj, S. Gollamudi, S. Kapoor, Y. F. Huang and J. R. Deller Jr., “Adaptive Interference Suppression for CDMA Systems with a WorstCase Error Criterion,” IEEE Transactions on Signal Processing, vol. 48, No. 1, January 2000. S. Gollamudi, S. Nagaraj, and Y. F. Huang, “Blind Equalization with a Deterministic Constant Modulus Cost - A Set-Membership Filtering Approach”, Proc. IEEE ICASSP, 2000. I. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci, ”A survey on sensor networks,” IEEE Communications Magazine, vol. 40, Issue 8, August, 2002, pp. 102-114. T. Wang, R. C. de Lamare, and P. D. Mitchell, ”Low-Complexity Set-Membership Channel Estimation for Cooperative Wireless Sensor Networks,” IEEE Transactions on Vehicular Technology, vol.60, no.6, pp.2594-2607, July 2011. D. Maksarov and J. P. Norton, “Tuning of noise bounds in parameter set estimation,” in Proceedings of International Conference on Identification in Engineering Systems, Swansea, UK, 27-29, March 1996, pp. 584-593. S. Gazor and K. Shahtalebi, “A New NLMS Algorithm for Slow Noise Magnitude Variation”, IEEE Signal Processing Letters, vol. 9, no. 11, November 2002. L. Guo and Y. F. Huang, “Set-Membership Adaptive Filtering with Parameter-Dependent Error Bound Tuning,” IEEE Proc. Int. Conf. Acoust. Speech and Sig. Proc., 2005. L. Guo and Y. F. Huang, “Frequency-Domain Set-Membership Filtering and Its Applications”, IEEE Transactions on Signal Processing, Vol. 55, No. 4, April 2007, pp. 1326-1338. M. Z. A. Bhotto, and A. Antoniou, “Robust Set-Membership AffineProjection Adaptive-Filtering Algorithm”, IEEE Transactions on Signal Processing, vol. 60 , no. 1, January 2012 , pp. 73 - 81. T. M. Lin, M. Nayeri and J. R. Deller, Jr., “Consistently Convergent OBE Algorithm with Automatic Selection of Error Bounds,” International Journal of Adaptive Control and Signal Processing, vol. 12, pp. 302-324, June 1998. J. R. Deller, S. Gollamudi, S. Nagaraj, and Y. F. Huang, “Convergence analysis of the QUASI-OBE algorithm and the performance implications,” in Proc. IFAC Int. Symp. Syst. Ident., vol. 3, Santa Barbara, CA, June 2000, pp. 875-880. S. Dasgupta and Y. F. Huang, “Asymptotically convergent modified recursive least-squares with data dependent updating and forgetting factor for systems with bounded noise,” IEEE Trans. Inform. Theory, vol. IT-33, pp. 383–392, 1987. J. R. Deller, M. Nayeri, and M. Liu, “Unifying the landmark developments in optimal bounding ellipsoid identification,” Int. J. Adap. Contr. Signal Process., vol. 8, pp. 43–60, 1994. D. Joachim, J. R. Deller, and M. Nayeri, “Multiweight optimization in OBE algorithms for improved tracking and adaptive identification,” in Proc. IEEE Int. Acoust., Speech Signal Process., vol. 4, Seattle, WA, May 1998, pp. 2201-2204. R. C. de Lamare and P. S. R. Diniz, “Set-Membership Adaptive Algorithms based on Time-Varying Error Bounds for CDMA Interference Suppression”, IEEE Transactions on Vehicular Technology, vol. 58, no. 2, February 2009 , pp. 644 - 654.

14

[22]

[23] [24] [25]

[26] [27]

[28] [29]

[30] [31] [32] [33]

[34]

[35]

[36]

[37]

[38] [39] [40] [41] [42] [43] [44]

R. C. de Lamare and P. S. R. Diniz, ”Blind constrained set-membership algorithms with time varying bounds for CDMA Interference Suppression”, in Proc. IEEE International Conference on Acoustics, Speech and Signal Processing, May 2006. J.C. Liberti and T. S. Rappaport, Smart Antennas for Wireless Communications: IS-95 and Third Generation CDMA Applications, Prentice Hall: Upper Saddle River, New Jersey, 1999. M. Honig, U. Madhow and S. Verdu, “Blind adaptive multiuser detection,” IEEE Trans. Inf. Theory, vol. 41, pp. 944-960, July 1995. R. C. de Lamare and R. Sampaio-Neto, “Adaptive Reduced-Rank Processing Based on Joint and Iterative Interpolation, Decimation, and Filtering,” IEEE Transactions on Signal Processing, vol. 57, no. 7, July 2009, pp. 2503 - 2514. Z. Xu and M.K. Tsatsanis, “Blind adaptive algorithms for minimum variance CDMA receivers,” IEEE Trans. Communications, vol. 49, No. 1, January 2001. R. C. de Lamare and R. Sampaio-Neto, “Low-Complexity Variable StepSize Mechanisms for Stochastic Gradient Algorithms in Minimum Variance CDMA Receivers”, IEEE Trans. Signal Processing, vol. 54, pp. 2302 - 2317, June 2006. R. C. de Lamare, “Adaptive Reduced-Rank LCMV Beamforming Algorithms Based on Joint Iterative Optimisation of Filters”, Electronics Letters, vol. 44, no. 9, 2008. R. Fa, R. C. de Lamare and L. Wang, “Reduced-rank STAP schemes for airborne radar based on switched joint interpolation, decimation and filtering algorithm”, IEEE Trans. Sig. Proc., 2010, vol. 58, no. 8, pp.41824194. J. Miguez and L. Castedo, “A linearly constrained constant modulus approach to blind adaptive multiuser interference suppression,” IEEE Communications Letters, vol. 2, pp. 217-219, August 1998. C. Xu, G. Feng and K. S. Kwak, “A Modified Constrained Constant Modulus Approach to Blind Adaptive Multiuser Detection,” IEEE Trans. Communications, vol. 49, No. 9, 2001. Z. Xu and P. Liu, “Code-Constrained Blind Detection of CDMA Signals in Multipath Channels,” IEEE Sig. Proc. Letters, vol. 9, No. 12, December 2002. R. C. de Lamare and R. Sampaio Neto, ”Blind Adaptive CodeConstrained Constant Modulus Algorithms for CDMA Interference Suppression in Multipath Channels”, IEEE Communications Letters, vol 9. no. 4, April, 2005. R. C. de Lamare, M. Haardt and R. Sampaio-Neto, “Blind Adaptive Constrained Reduced-Rank Parameter Estimation based on Constant Modulus Design for CDMA Interference Suppression,” IEEE Transactions on Signal Processing, vol. 56., no. 6, June 2008. Y. Cai and R. C. de Lamare, “Low-Complexity Variable Step Size Mechanism for Code-Constrained Constant Modulus Stochastic Gradient Algorithms applied to CDMA Interference Suppression”, IEEE Transactions on Signal Processing, vol. 57, No. 1, January 2009. R. C. de Lamare and R. Sampaio-Neto, “Blind adaptive MIMO receivers for space-time block-coded DS-CDMA systems in multipath channels using the constant modulus criterion,” IEEE Transactions on Communications, vol. 58, no. 1, January 2010, pp. 21-27. R.C. de Lamare, R. Sampaio-Neto and M. Haardt, ”Blind Adaptive Constrained Constant-Modulus Reduced-Rank Interference Suppression Algorithms Based on Interpolation and Switched Decimation,” IEEE Trans. on Signal Processing, vol.59, no.2, pp.681-695, Feb. 2011. J. Mai and A. H. Sayed,“A feedback approach to the steady-state performance of fractionally spaced blind adaptive equalizers,” IEEE Trans. Signal Process., vol. 48, no. 1, pp. 80–91, Jan. 2000. N. R. Yousef and A. H. Sayed, “A unified approach to the steady-state and tracking analyses of adaptive filters,” IEEE Trans. Signal Process., vol. 49, no. 2, pp. 314–324, Feb. 2001. J. Whitehead and F. Takawira, “Performance analysis of the linearly constrained constant modulus algorithm-based multiuser detector,” IEEE Trans. Signal Process., vol. 53, no. 2, pp. 643–653, Feb. 2005. X. G. Doukopoulos and G. V. Moustakides, “Adaptive Power Techniques for Blind Channel Estimation in CDMA Systems”, IEEE Trans. Signal Processing, vol. 53, No. 3, March, 2005. X. Wang and H. V. Poor, “Blind multiuser detection: A subspace approach,” IEEE Transactions on Information Theory, vol. 44, pp. 677690, March 1998. D. P. Bertsekas, Nonlinear Programming, Athena Scientific, 2nd Ed., 1999. T. S. Rappaport, Wireless Communications, Prentice-Hall, Englewood Cliffs, NJ, 1996.